Parameter Tuning of Linear Programming Solvers

Abstract

Linear programming solvers include various options that can be used to control algorithmic aspects and considerably impact the solver performance. As it is obvious, manually finding optimal parameters is a very difficult task and sometimes impossible. For this reason, it is necessary to implement smart techniques that will automate this process. Other works have utilized derivative-free optimization solvers to tune solver parameters. In this work, eight open-source derivative-free optimization solvers are utilized for finding (near) optimal tuning parameters of state-of-the-art linear programming solvers. We investigate how sensitive linear programming solvers are to a parameter tuning process. Extensive computational results are presented on tuning four linear programming solvers (CLP, CPLEX, GUROBI, and XPRESS) over a set of 70 benchmark problems. We find better parameters for all linear programming solvers, achieving a reduction in execution time over their default parameters up to 26%. We conclude that several derivative-free optimization solvers outperform others on finding optimal optimal tuning parameters for linear programming solvers.